JP7383684B2 - Information processing device, information processing method, and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program, and particularly to an information processing device, an information processing method, and a program for relearning image classification AI.

Conventionally, in order to generate highly accurate image classification AI, it has been necessary to repeatedly retrain the image classification AI using input images as training data and then verify the model generated by the relearning. It is being said. In addition, when relearning, it is necessary to identify and prepare learning data necessary to improve accuracy based on the verification results, and existing images are sometimes processed to be used as learning data.

As a method of processing input images used as learning data to improve the performance of image classification AI, a method of replacing a part of the input image with other values so as to hide it has been proposed.

Non-Patent Document 1 describes replacing random positions in an input image with rectangles having random sizes and values.

Non-Patent Document 2 describes replacing square areas with 0 centered at random positions in an input image.

Z. Zhong, L. Zheng, G. Kang, S. Li, and Y. Yang, “Random Erasing Data Augmentation”, in arXiv:1708.04896, 2017 T. Devries and G. W. Taylor. “Improved regulation of convolutional neural networks with cutout.” arXiv preprint arXiv: 1708.04552, 2017.5

In the techniques described in Non-Patent Documents 1 and 2, the processing positions applied to the input image used as learning data are random, so even if the area in the input image that is likely to contribute to improving the performance of image classification AI is known, the processing position is random. It is not possible to specify an area and perform pinpoint processing.

Therefore, an object of the present invention is to provide an information processing device, an information processing method, and a program that can efficiently generate learning data that contributes to improving the accuracy of image classification AI.

In order to solve the above problems, an information processing apparatus according to the present invention includes a processing reception means for receiving an image to be processed from among a plurality of input images to be processed for outputting a predicted label by a trained model ; an acquisition means for acquiring a feature region that is the basis for prediction of the learned model; and an image accepted as an image to be processed by the processing reception means , and selected with a first probability ; Performing a first processing process to replace the values of at least some pixels in the characteristic region with a first predetermined value, and accepting the image as a target image to be processed by the processing receiving means among the plurality of input images. processing means for performing a second processing process for replacing the values of some pixels of the image with a second predetermined value on an image that is not selected with a second probability; The present invention is characterized by comprising an update unit that updates learning data for relearning a trained model using images subjected to the first and second processing.

According to the present invention, learning data that contributes to improving the accuracy of image classification AI can be efficiently generated.

1 is a diagram illustrating an example of a system configuration of an AI prediction basis display system including an information processing device according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example of a hardware configuration of an information processing device. It is a flow chart which shows an example of relearning/evaluation processing concerning an embodiment of the present invention. This is a result analysis screen displayed on the display device in step S301 of FIG. 3. This is a learning table in which paths of input images to be added to learning data are registered. This is a processing table in which paths of input images to be added to or replaced with learning data after processing are registered. This is a deletion table in which paths of input images to be deleted from evaluation data or learning data are registered. 5 is a flowchart illustrating an example of data processing processing executed in steps S411 and S412 of FIG. 4. FIG. 5 is a flowchart showing another example of data processing processing executed in steps S411 and S412 of FIG. 4. FIG. FIG. 3 is a diagram showing an example of an input image, a prediction basis image, and a processed image.

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a diagram showing an example of the system configuration of an AI prediction basis display system 1 including an information processing device 101 according to an embodiment of the present invention.

The AI prediction basis display system 1 is a system that includes an information processing device 101 and an external device 102, which are communicably connected via a network 110.

The information processing device 101 is a device operated by a user, and displays input images to be processed in relearning/evaluation processing (FIG. 3), which will be described later, and data linked thereto on a result analysis screen (FIG. 4). At the same time, presses of various buttons on this screen by the user are accepted.

The external device 102 manages input images to be processed in the relearning/evaluation process and data associated therewith.

In the embodiment of the present invention, the information processing apparatus 101 will be described as executing the process shown in the flowchart of FIG. 3, but the external apparatus 102 may execute the process. Further, the information processing apparatus 101 may perform processing described as being performed by the external device 102, such as management of images to be processed.

FIG. 2 is a block diagram showing an example of the hardware configuration of the information processing apparatus 101. Note that since the external device 102 has the same hardware configuration as the information processing device 101, repeated explanation will be omitted.

In FIG. 2, the information processing device 101 includes a CPU 201, a ROM 202, a RAM 203, a storage device 204, an input control section 205, an audio control section 206, a video control section 207, a memory control section 208, and a communication I/F control section 209. . These devices and controllers are connected to each other via a system bus 200.

The CPU 201 centrally controls each device and controller connected to the system bus 200.

The RAM 203 functions as the main memory, work area, etc. of the CPU 201. The CPU 201 loads necessary programs and the like from the ROM 202 or the external memory 213 into the RAM 203 when executing processing, and executes the loaded programs to realize various operations.

The storage device 204 is comprised of an SSD, an HDD, or the like, and stores image classification AI, etc., which will be described later.

The input control unit 205 controls input from an input device 210 including a keyboard, a touch panel, a pointing device such as a mouse, and the like. For example, if the input device 210 is a touch panel, the user can issue various instructions by pressing (touching with a finger or the like) an icon, cursor, or button displayed on the touch panel. Note that the touch panel in this case may be a touch panel capable of detecting positions touched by multiple fingers, such as a multi-touch screen.

The audio control unit 206 controls audio input/output to the audio input/output device 211 such as a microphone and a speaker.

A video control unit 207 controls display on a display device 212 such as a display or a projector. In this case, the display includes the display of a notebook computer that is integrated with the main body. Note that if the display device 212 is a device that can accept the above-mentioned touch operation, it also serves as the input device 210. Further, the video control unit 207 can control a video memory (VRAM) for display control, and may use a part of the RAM 203 as the video memory area, or may use a separate dedicated video memory. A memory may also be provided.

Memory control unit 208 controls access to external memory 213. The external memory 213 is not particularly limited as long as it is a storage device that stores boot programs, various applications, font data, user files, editing files, various data, and the like. For example, an external storage device (hard disk), a flexible disk (FD), a Compact Flash (registered trademark) memory connected to a PCMCIA card slot via an adapter, etc. can be used as the external memory 213.

The communication I/F control unit 209 connects and communicates with external devices via the network 110, and executes communication control processing on the network 110. For example, communication using TCP/IP, a telephone line such as ISDN, a 4G line, a 5G line of a mobile phone, etc. is possible.

Note that the CPU 201 enables display on the display device 212 by, for example, executing an outline font development (rasterization) process in a display information area in the RAM 203. Furthermore, the CPU 201 can also accept user instructions using a mouse cursor (not shown) on the display device 212.

Next, the relearning/evaluation process according to the embodiment of the present invention will be described using the flowchart in FIG. 3.

This process is executed by the CPU 201 of the information processing device 101 reading a predetermined control program. In this process, information regarding the prediction results of the image classification AI is displayed on the result analysis screen (FIG. 4). In addition, after the learning data and evaluation data are updated according to the operations received from the user on the result analysis screen (pressing the learning/processing/delete button), image classification is performed using the updated learning data and evaluation data. AI will be retrained and evaluated. The learning data and evaluation data will be described later.

First, in step S<b>301 , the CPU 201 acquires an input image and data associated therewith, specifically, correct label, predicted label, and heat map information, from the external device 102 and stores them in the RAM 203 . Thereafter, the CPU 201 controls to display the correct label, predicted label, input image, and prediction basis image on the result analysis screen (FIG. 4: display means) of the display device 212 based on the input image and data associated therewith. do.

Here, the correct label is a label that is predetermined by the user or the like and indicates the correct answer for the inference made by the image classification AI on the input image.

The image classification AI is a trained model using a Convolutional Neural Network (CNN) that is trained in advance by a learning unit (not shown in FIG. 1) in the external device 102 and stored in the storage device 204. Although CNN is used for image classification AI here, the invention is not limited to this, and other known algorithms such as support vector machine may also be used.

When the CPU 201 receives an input image (and its correct label) from the outside, it reads the image classification AI from the storage device 204 and inputs the received input image into the read image classification AI, thereby generating a prediction inferred from the input image. Output the label. The CPU 201 transmits a plurality of input images and a correct label and a predicted label as data associated with each of these to the external device 102 before this processing.

The heat map is a feature region in the input image that is the basis for the prediction of the image classification AI, and the CPU 201 calculates the heat map using Grad-CAM, which is a known technology. The CPU 201 transmits this heat map information to the external device 102 in advance as information linking each of the plurality of input images before this processing. Although Grad-CAM is used to calculate the heat map here, the calculation is not limited to this, and other known algorithms such as RISE, Score-CAM, XRAI, Group-CAM, etc. may also be used.

In this way, the CPU 201 stores the input image and the correct label, predicted label, and heat map information associated therewith in the external device 102 in advance before this processing. Input images are managed so that they can be identified as learning data or evaluation data, such as by linking and registering information indicating whether the input images are learning data or evaluation data. shall be taken as a thing. Further, in this embodiment, the CPU 201 calculates the heat map, but the present invention is not limited to this, as long as the heat map information is stored in the external device 102 in advance before this processing. For example, the CPU 201 may acquire a heat map calculated externally.

The prediction basis image is an image in which a heat map is superimposed on the input image to visualize the prediction basis of the image classification AI. The heat map in the prediction basis image is colored. Specifically, the higher the degree of prediction basis, the warmer color is assigned to the characteristic region of the heat map, and the lower the degree of prediction basis, the colder color is assigned to the characteristic region of the heat map. For example, the feature regions of the heat map are colored red, orange, yellow, yellow-green, and light blue in descending order of the degree of basis for prediction.

In step S302, the CPU 201 determines whether the input image (target) received in step S301 is evaluation data. As a result of this determination, if the target is evaluation data (YES in step S302), the process advances to step S303; otherwise, the process advances to step S305.

In step S303, the CPU 201 determines whether the learning button 405 on the result analysis screen in FIG. 4 has been pressed. As a result of this determination, if the learning button 405 has been pressed (YES in step S303), the process advances to step S304; otherwise, the process advances to step S305.

In step S304, the CPU 201 registers the path of the target input image in the learning table (FIG. 5) in order to improve the accuracy of the image classification AI by performing re-learning using the target input image and its correct answer label. do. For example, if the predicted label of a target is incorrect, retraining the image classification AI using the target input image and its correct label can be expected to prevent similar mistakes from being made. Compared to the case of data processing described later, it is possible to correct the predicted label output by image classification AI regardless of the original prediction basis, so it is considered suitable when the prediction basis is correct but the predicted label is wrong. It will be done.

In step S305, the CPU 201 determines whether the processing button 406 on the result analysis screen in FIG. 4 has been pressed. As a result of this determination, if the processing button 406 has been pressed (YES in step S305), the process advances to step S306; otherwise, the process advances to step S307.

In step S306, the CPU 201 (registration means) registers the path of the target input image in the processing table (FIG. 6) in order to perform data processing. Note that the data processing process refers to a process that performs processing such as replacing the inappropriate part in the input image with a single value when the inappropriate part becomes a feature part that is the basis for prediction.

Here, the data processing process illustrated in FIG. 8, which will be described later, is a process that is assumed to be performed on all input images whose paths are registered in the processing table when the export button 408 (FIG. 4) is pressed. be. In addition, the data processing process illustrated in FIG. 9, which will be described later, allows the probability of processing to be set in order to prevent the consistency of the processing itself from being used as a basis for prediction. This is a process that is assumed to be performed on an input image. In this way, if the prediction basis is inappropriate, by processing the input image and relearning the image classification AI using this, the image classification AI will learn the correct prediction basis and make the correct prediction. can be expected to do so. Compared to the above-mentioned relearning process, the data processing process can correct the prediction basis and is therefore considered suitable for cases where the prediction basis is incorrect and the predicted label is also incorrect.

In step S307, the CPU 201 determines whether the delete button 407 on the result analysis screen in FIG. 4 has been pressed. As a result of this determination, if the delete button 407 has been pressed (YES in step S307), the process advances to step S308; otherwise, the process advances to step S309. Here, the delete button 407 is used to delete the target input image from the learning data and evaluation data because if the user uses the target input image as training data and evaluation data, it will have a negative effect on the relearning and evaluation of the image classification AI. This button is pressed when you want to perform deletion processing. Here, cases in which the target input image has a negative effect on the relearning and evaluation of image classification AI include, for example, cases in which the input image included in the input information is an image unrelated to learning or evaluation, or the case where the target input image This is the case when the image is too large or too small.

In step S308, the CPU 201 registers the path of the target input image in the deletion table (FIG. 7) in order to execute the deletion process.

In step S309, the CPU 201 determines whether the export button 408 (update accepting means) on the result analysis screen in FIG. 4 has been pressed. As a result of this determination, if the export button 408 has been pressed (YES in step S309), the process advances to step S310, and if not, the process returns to step S301.

In step S310, the CPU 201 determines whether the target input image is evaluation data. As a result of this determination, if the target is evaluation data (YES in step S310), the process advances to step S311, and if not, that is, if the target is learning data (NO in step S310), the process advances to step S312.

In step S311, if the path of the target input image is registered in the learning table, the CPU 201 (updating unit) changes the target input image from evaluation data to learning data. That is, when the learning result display unit 400 receives the pressing of the export button 408 after pressing the learning button 405, the storage destination of the target input image is moved from the evaluation data folder of the external device 102 to the learning data folder. Further, if the path of the target input image is registered in the processing table, the CPU 201 processes the target input image through data processing processing, and changes the processed input image from evaluation data to learning data. That is, when the learning result display unit 400 receives the press of the export button 408 after pressing the processing button 406, the storage destination of the target input image after the processing is changed from the evaluation data folder of the external device 102 to the learning data folder. Move to. Furthermore, if the path of the target input image is registered in the deletion table, the CPU 201 deletes the target input image from the evaluation data. That is, when the learning result display unit 400 receives the press of the export button 408 after pressing the delete button 407, the target input image is deleted from the evaluation data folder of the external device 102 where it is currently stored. After that, the process advances to step S313.

In this way, when the target is evaluation data, not only the evaluation data but also the learning data is updated. In this case, if the export button 408 is pressed after only the delete button 407 is pressed, the learning data will not be updated, so there is no need to re-learn the image classification AI in step S313, which will be described later.

In step S312, if the path of the target input image is registered in the processing table, the CPU 201 (updating means) processes the target input image through data processing processing, and uses the processed input image as learning data. Replace the input image for the target. That is, the target input image in the learning data folder of the external device 102 is replaced with the target input image after the above processing. Further, if the path of the target input image is registered in the deletion table, the CPU 201 deletes the target input image from the learning data. That is, the target input image is deleted from the learning data folder of the external device 102 where it is currently stored. After that, the process advances to step S313.

In this way, when the target is learning data, only the learning data is updated, and the evaluation data is not updated. Therefore, when the target is learning data, it is not necessary to evaluate the image classification AI in step S313, which will be described later.

In step S313, the CPU 201 outputs learning data from the learning data folder of the external device 102 and re-learns the image classification AI. Further, the CPU 201 outputs evaluation data from the evaluation data folder of the external device 102 and performs image classification AI evaluation. After that, this process ends.

In this embodiment, the CPU 201 of the information processing apparatus 101 performs relearning of the image classification AI, but if the relearning of the image classification AI is performed using updated learning data, this embodiment may be used. Not limited. For example, the CPU 201 may instruct the external device 102 to perform relearning of the image classification AI, and obtain the relearned image classification AI (or its internal parameters) from the external device 102. Similarly, although the CPU 201 of the information processing apparatus 101 executes the evaluation of the image classification AI, the present invention is not limited to this embodiment as long as the evaluation of the image classification AI is performed using updated evaluation data. For example, the CPU 201 may instruct the external device 102 to perform image classification AI evaluation, and obtain the evaluation results from the external device 102.

FIG. 4 is a result analysis screen displayed on the display device 212 in step S301 of FIG.

The result analysis screen includes learning result display sections 400a to 400d (hereinafter collectively referred to as "learning result display sections 400") for each input image, and an export button 408.

The learning result display section 400 includes a correct label display section 401, a predicted label display section 402, an input image display section 403, a prediction basis image display section 404, a learning button 405, a processing button 406, and a deletion button 407.

The correct label display section 401 displays the correct label of the input image.

The predicted label display unit 402 displays the predicted label of the image classification AI.

Input image display section 403 displays an input image.

A prediction basis image display section 404 displays a prediction basis image.

In this way, the correct label, predicted label, input image, and prediction basis image are displayed in a comparable manner on one screen, the result analysis screen in FIG. 4. This allows the user to easily judge the validity of the learning results using the result analysis screen shown in FIG.

The learning button 405 is a button that accepts an instruction to register the path of the input image in the learning table.

The processing button 406 (processing acceptance means) is a button that accepts an instruction (registration instruction) to register the path of the input image in the processing table.

The delete button 407 is a button that accepts an instruction to register the path of the input image in the deletion table.

The export button 408 is a button that accepts an instruction (update instruction) to process input images registered in the learning table, processing table, and deletion table, and output the updated learning data and/or evaluation data.

If the user determines that the learning results displayed on the learning result display section 400 are not valid, the user presses at least one of the learning button 405, processing button 406, and delete button 407 on the result analysis screen in FIG. Press button 408. By only operating the learning result display section 400, the user can easily implement measures to improve the accuracy of image classification AI. As specific accuracy improvement measures, it is possible to add the input image as learning data, process the input image as learning data and evaluation data, and delete the input image from the learning data and evaluation data.

FIG. 5 is a learning table in which paths of input images to be added to learning data are registered.

FIG. 6 is a processing table in which paths of input images to be added to or replaced with learning data after processing are registered.

FIG. 7 is a deletion table in which paths of input images to be deleted from evaluation data or learning data are registered.

Next, an example of the data processing processing executed in steps S411 and S412 in FIG. 4 will be described using the flowchart in FIG. 8. This process is executed for all input images whose paths are registered in the processing table when the export button 408 (FIG. 4) is pressed. That is, in this process, all input images whose paths are registered in the processing table are processed so that the image classification AI is not trained based on incorrect prediction grounds.

This process is executed by the CPU 201 of the information processing device 101 reading a predetermined control program.

First, in step S801, the CPU 201 determines one of the input images whose path is registered in the processing table to be processed. An example of the input image determined to be processed here is shown in FIG. 10(A). In this input image, the position of the left hole is shifted from the designed position, so the correct label is "abnormal." Therefore, it is desirable that the predicted label inferred from this input image by the image classification AI is also "abnormal". Further, it is desirable that the heat map of the prediction basis image be given a warmer color (a color indicating a higher degree of prediction basis) closer to the position of the hole on the left.

In step S802, the CPU 201 acquires the coordinates of pixels in the heat map that have a value larger than the set value 1A (predetermined standard). Each pixel of the heatmap takes a value between 0 and 1. Further, the setting value 1A is a threshold value for determining whether or not the degree of prediction based on each pixel of the heat map satisfies a predetermined criterion, and has a value greater than 0 and less than 1. An example of the prediction basis image is shown in FIG. 10(B). In this prediction basis image, the area around the hole on the right is used as the basis for prediction, and it can be seen that the prediction label was output based on an incorrect prediction basis. In step S802, such an image is processed.

In step S803, the CPU 201 processes the input image to replace the values of pixels at all coordinates acquired in step S802 with a set value 2A (first predetermined value). FIG. 10(C) shows an image (processed image) obtained by processing the input image of FIG. 10(a) in step S803. In this processed image, processing is performed with black specified as the setting value 2A.

In step S804, the CPU 201 determines whether the processing in step S803 has been completed for all input images whose paths are registered in the processing table. As a result of the determination, if the processing in step S803 has been completed for all input images (YES in step S804), this process is ended, whereas if not (NO in step S804), the process returns to step S801.

Next, another example of the data processing processing executed in steps S411 and S412 of FIG. 4 will be described using the flowchart of FIG. 9. In this process, the input image is processed with a certain probability in order to prevent the consistency of the input image processing itself from affecting heat map calculation. Furthermore, when the export button 408 (FIG. 4) is pressed, this processing is performed not only for input images whose paths are registered in the processing table but also for all input images saved in the external device 102. The processing is performed stochastically. This process is intended to be applied especially during additional learning. That is, in this process, processing is applied to all input images in the information processing apparatus 101 in a probabilistic manner so that the image classification AI is not trained based on incorrect prediction grounds.

This process is executed by the CPU 201 of the information processing device 101 reading a predetermined control program.

First, in step S901, the CPU 201 acquires one of the input images stored in the external device 102.

In step S902, the CPU 201 determines whether the path of the input image acquired in step S901 is registered in the processing table. As a result of the determination, if it is registered in the processing table (YES in step S902), the process advances to step S903, and if not (NO in step S902), the process advances to step S906.

In step S903, the CPU 201 generates a random number and determines whether the generated random number is larger than the set value 1B (first probability). The random number generated here and the set value 1B both take values between 0 and 1. As a result of the determination, if the random number is larger than the set value 1B (YES in step S903), the process advances to step S904 to process the input image acquired in step S901; otherwise (NO in step S903), the process proceeds to step S904. Finish the process. In this way, applying the processing stochastically is expected to play a role in preventing the consistency of the processing itself from affecting the calculation of the heat map.

In step S904, the CPU 201 acquires the coordinates of pixels in the heat map that have values greater than the set value 2B (predetermined standard), and proceeds to step S905. Each pixel of the heatmap takes a value between 0 and 1. Further, the setting value 2B is a threshold value for determining whether or not the degree of prediction based on each pixel of the heat map satisfies a predetermined criterion, and has a value greater than 0 and less than 1.

In step S905, the CPU 201 processes the input image obtained in step S901 to replace the values of pixels at all coordinates obtained in step S904 with the set value 3B (first predetermined value), and then performs the main processing. end.

In step S906, the CPU 201 generates a random number and determines whether the generated random number is larger than the set value 1B (second probability). As a result of the determination, if the random number is larger than the set value 1B (YES in step S906), the process advances to step S907 to process the input image acquired in step S901; otherwise (NO in step S906), the process proceeds to step S907. Finish the process. In this way, by probabilistically processing input images that are stored in the external device 102 and whose paths are not registered in the processing table, the consistency of the processing itself can be maintained in the heat map calculation. It is expected that it will play a role in preventing negative impacts. Although the setting value used for the determination in step S906 is the same as the setting value 1B used in the determination in step S903, it may be a different setting value.

In step S907, the CPU 201 randomly selects a pixel in the input image acquired in step S901, replaces the value of the randomly selected pixel with the setting value 3B (second predetermined value), and then starts this process. finish. Note that the number of pixels randomly selected here may also be a randomly set value or may be a fixed value.

(Other embodiments)
In this embodiment, a program for realizing one or more functions is supplied to a computer of a system or device via a network or a storage medium, and a system control unit of the system or device reads and executes the program. It is possible. The system controller has one or more processors or circuits and may include separate system controllers or a network of separate processors or circuits for reading and executing executable instructions.

The processor or circuit may include a central processing unit (CPU), microprocessing unit (MPU), graphics processing unit (GPU), application specific integrated circuit (ASIC), field programmable gate array (FPGA). The processor or circuit may also include a digital signal processor (DSP), a data flow processor (DFP), or a neural processing unit (NPU).

Although preferred embodiments of the present invention have been described above, the present invention is not limited to these embodiments, and various modifications and changes can be made within the scope of the invention.

1 AI prediction basis display system 101 Information processing device 102 External device 110 Network 201 CPU
212 Display device 405 Learning button 406 Processing button 407 Delete button 408 Export button

Claims (7)

processing reception means for receiving an image to be processed from among a plurality of input images to be processed for outputting predicted labels by the learned model;
acquisition means for acquiring a feature region that is the basis for prediction of the learned model;
For an image accepted as an image to be processed by the processing receiving means and selected with a first probability , the values of at least some pixels of the characteristic region are set to a first predetermined value. A first processing process is performed to replace the image with a second probability, and among the plurality of input images, the image is not accepted as an image to be processed by the processing acceptance means, and is selected with a second probability. processing means for performing a second processing process to replace the values of some pixels of the image with a second predetermined value;
updating means for updating learning data for relearning the trained model using the images subjected to the first and second processing;
An information processing device comprising: The second processing is performed on images that are not accepted as processing targets by the processing acceptance means among the plurality of input images and that are selected with the second probability. The information processing apparatus according to claim 1, further comprising a process of replacing a value of a pixel in the selected area with the second predetermined value. registration means for registering, among the plurality of input images, an image for which the processing reception means has received a registration instruction as an image to be processed;
update receiving means for receiving an update instruction to update learning data for relearning the learned model;
Equipped with
When the update receiving means receives the update instruction, the updating means performs the first and second processing by the processing means , and uses the image on which the first and second processing has been performed. The information processing device according to claim 1 or 2, wherein the learning data is updated by using the learning data. the plurality of input images, the prediction label, a prediction basis image obtained by superimposing the feature region on the input image , and a correct answer label indicating the correct answer of the inference when inputting the predetermined input image to the learned model. 4. The information processing apparatus according to claim 1, further comprising display means for displaying the information on one screen in a comparative manner. 5. At least some of the pixels in the feature region are pixels located in a region of the feature region whose degree of prediction basis satisfies a predetermined criterion. The information processing device described in . An information processing method for an information processing device, the method comprising:
a processing reception step of receiving an image to be processed from among a plurality of input images to be processed to output predicted labels by the learned model;
an acquisition step of acquiring a feature region that is the basis for prediction of the learned model;
For an image accepted as an image to be processed in the processing reception step and selected with a first probability , the values of at least some pixels of the characteristic region are set to a first predetermined value. A first processing process is performed to replace the image with a second probability , and among the plurality of input images, the image is not accepted as an image to be processed in the processing acceptance step, and is selected with a second probability. a processing step of performing a second processing process of replacing the values of some pixels of the image with a second predetermined value;
an updating step of updating learning data for relearning the trained model using the images subjected to the first and second processing;
An information processing method comprising: A computer-executable program that causes a computer to function as each means of the information processing apparatus according to claim 1 .

Images